include/uapi/linux/perf_event.h

   1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
   2 /*
   3  * Performance events:
   4  *
   5  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   7  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   8  *
   9  * Data type definitions, declarations, prototypes.
  10  *
  11  *    Started by: Thomas Gleixner and Ingo Molnar
  12  *
  13  * For licencing details see kernel-base/COPYING
  14  */
  15 #ifndef _UAPI_LINUX_PERF_EVENT_H
  16 #define _UAPI_LINUX_PERF_EVENT_H
  17
  18 #include <linux/types.h>
  19 #include <linux/ioctl.h>
  20 #include <asm/byteorder.h>
  21
  22 /*
  23  * User-space ABI bits:
  24  */
  25
  26 /*
  27  * attr.type
  28  */
  29 enum perf_type_id {
  30         PERF_TYPE_HARDWARE                      = 0,
  31         PERF_TYPE_SOFTWARE                      = 1,
  32         PERF_TYPE_TRACEPOINT                    = 2,
  33         PERF_TYPE_HW_CACHE                      = 3,
  34         PERF_TYPE_RAW                           = 4,
  35         PERF_TYPE_BREAKPOINT                    = 5,
  36
  37         PERF_TYPE_MAX,                          /* non-ABI */
  38 };
  39
  40 /*
  41  * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE
  42  * PERF_TYPE_HARDWARE:                  0xEEEEEEEE000000AA
  43  *                                      AA: hardware event ID
  44  *                                      EEEEEEEE: PMU type ID
  45  * PERF_TYPE_HW_CACHE:                  0xEEEEEEEE00DDCCBB
  46  *                                      BB: hardware cache ID
  47  *                                      CC: hardware cache op ID
  48  *                                      DD: hardware cache op result ID
  49  *                                      EEEEEEEE: PMU type ID
  50  * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied.
  51  */
  52 #define PERF_PMU_TYPE_SHIFT             32
  53 #define PERF_HW_EVENT_MASK              0xffffffff
  54
  55 /*
  56  * Generalized performance event event_id types, used by the
  57  * attr.event_id parameter of the sys_perf_event_open()
  58  * syscall:
  59  */
  60 enum perf_hw_id {
  61         /*
  62          * Common hardware events, generalized by the kernel:
  63          */
  64         PERF_COUNT_HW_CPU_CYCLES                = 0,
  65         PERF_COUNT_HW_INSTRUCTIONS              = 1,
  66         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
  67         PERF_COUNT_HW_CACHE_MISSES              = 3,
  68         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
  69         PERF_COUNT_HW_BRANCH_MISSES             = 5,
  70         PERF_COUNT_HW_BUS_CYCLES                = 6,
  71         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
  72         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
  73         PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
  74
  75         PERF_COUNT_HW_MAX,                      /* non-ABI */
  76 };
  77
  78 /*
  79  * Generalized hardware cache events:
  80  *
  81  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
  82  *       { read, write, prefetch } x
  83  *       { accesses, misses }
  84  */
  85 enum perf_hw_cache_id {
  86         PERF_COUNT_HW_CACHE_L1D                 = 0,
  87         PERF_COUNT_HW_CACHE_L1I                 = 1,
  88         PERF_COUNT_HW_CACHE_LL                  = 2,
  89         PERF_COUNT_HW_CACHE_DTLB                = 3,
  90         PERF_COUNT_HW_CACHE_ITLB                = 4,
  91         PERF_COUNT_HW_CACHE_BPU                 = 5,
  92         PERF_COUNT_HW_CACHE_NODE                = 6,
  93
  94         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
  95 };
  96
  97 enum perf_hw_cache_op_id {
  98         PERF_COUNT_HW_CACHE_OP_READ             = 0,
  99         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
 100         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
 101
 102         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
 103 };
 104
 105 enum perf_hw_cache_op_result_id {
 106         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
 107         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
 108
 109         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
 110 };
 111
 112 /*
 113  * Special "software" events provided by the kernel, even if the hardware
 114  * does not support performance events. These events measure various
 115  * physical and sw events of the kernel (and allow the profiling of them as
 116  * well):
 117  */
 118 enum perf_sw_ids {
 119         PERF_COUNT_SW_CPU_CLOCK                 = 0,
 120         PERF_COUNT_SW_TASK_CLOCK                = 1,
 121         PERF_COUNT_SW_PAGE_FAULTS               = 2,
 122         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
 123         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
 124         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
 125         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
 126         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
 127         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
 128         PERF_COUNT_SW_DUMMY                     = 9,
 129         PERF_COUNT_SW_BPF_OUTPUT                = 10,
 130         PERF_COUNT_SW_CGROUP_SWITCHES           = 11,
 131
 132         PERF_COUNT_SW_MAX,                      /* non-ABI */
 133 };
 134
 135 /*
 136  * Bits that can be set in attr.sample_type to request information
 137  * in the overflow packets.
 138  */
 139 enum perf_event_sample_format {
 140         PERF_SAMPLE_IP                          = 1U << 0,
 141         PERF_SAMPLE_TID                         = 1U << 1,
 142         PERF_SAMPLE_TIME                        = 1U << 2,
 143         PERF_SAMPLE_ADDR                        = 1U << 3,
 144         PERF_SAMPLE_READ                        = 1U << 4,
 145         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
 146         PERF_SAMPLE_ID                          = 1U << 6,
 147         PERF_SAMPLE_CPU                         = 1U << 7,
 148         PERF_SAMPLE_PERIOD                      = 1U << 8,
 149         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
 150         PERF_SAMPLE_RAW                         = 1U << 10,
 151         PERF_SAMPLE_BRANCH_STACK                = 1U << 11,
 152         PERF_SAMPLE_REGS_USER                   = 1U << 12,
 153         PERF_SAMPLE_STACK_USER                  = 1U << 13,
 154         PERF_SAMPLE_WEIGHT                      = 1U << 14,
 155         PERF_SAMPLE_DATA_SRC                    = 1U << 15,
 156         PERF_SAMPLE_IDENTIFIER                  = 1U << 16,
 157         PERF_SAMPLE_TRANSACTION                 = 1U << 17,
 158         PERF_SAMPLE_REGS_INTR                   = 1U << 18,
 159         PERF_SAMPLE_PHYS_ADDR                   = 1U << 19,
 160         PERF_SAMPLE_AUX                         = 1U << 20,
 161         PERF_SAMPLE_CGROUP                      = 1U << 21,
 162         PERF_SAMPLE_DATA_PAGE_SIZE              = 1U << 22,
 163         PERF_SAMPLE_CODE_PAGE_SIZE              = 1U << 23,
 164         PERF_SAMPLE_WEIGHT_STRUCT               = 1U << 24,
 165
 166         PERF_SAMPLE_MAX = 1U << 25,             /* non-ABI */
 167
 168         __PERF_SAMPLE_CALLCHAIN_EARLY           = 1ULL << 63, /* non-ABI; internal use */
 169 };
 170
 171 #define PERF_SAMPLE_WEIGHT_TYPE (PERF_SAMPLE_WEIGHT | PERF_SAMPLE_WEIGHT_STRUCT)
 172 /*
 173  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
 174  *
 175  * If the user does not pass priv level information via branch_sample_type,
 176  * the kernel uses the event's priv level. Branch and event priv levels do
 177  * not have to match. Branch priv level is checked for permissions.
 178  *
 179  * The branch types can be combined, however BRANCH_ANY covers all types
 180  * of branches and therefore it supersedes all the other types.
 181  */
 182 enum perf_branch_sample_type_shift {
 183         PERF_SAMPLE_BRANCH_USER_SHIFT           = 0, /* user branches */
 184         PERF_SAMPLE_BRANCH_KERNEL_SHIFT         = 1, /* kernel branches */
 185         PERF_SAMPLE_BRANCH_HV_SHIFT             = 2, /* hypervisor branches */
 186
 187         PERF_SAMPLE_BRANCH_ANY_SHIFT            = 3, /* any branch types */
 188         PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT       = 4, /* any call branch */
 189         PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT     = 5, /* any return branch */
 190         PERF_SAMPLE_BRANCH_IND_CALL_SHIFT       = 6, /* indirect calls */
 191         PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT       = 7, /* transaction aborts */
 192         PERF_SAMPLE_BRANCH_IN_TX_SHIFT          = 8, /* in transaction */
 193         PERF_SAMPLE_BRANCH_NO_TX_SHIFT          = 9, /* not in transaction */
 194         PERF_SAMPLE_BRANCH_COND_SHIFT           = 10, /* conditional branches */
 195
 196         PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT     = 11, /* call/ret stack */
 197         PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT       = 12, /* indirect jumps */
 198         PERF_SAMPLE_BRANCH_CALL_SHIFT           = 13, /* direct call */
 199
 200         PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT       = 14, /* no flags */
 201         PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT      = 15, /* no cycles */
 202
 203         PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT      = 16, /* save branch type */
 204
 205         PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT       = 17, /* save low level index of raw branch records */
 206
 207         PERF_SAMPLE_BRANCH_MAX_SHIFT            /* non-ABI */
 208 };
 209
 210 enum perf_branch_sample_type {
 211         PERF_SAMPLE_BRANCH_USER         = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
 212         PERF_SAMPLE_BRANCH_KERNEL       = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
 213         PERF_SAMPLE_BRANCH_HV           = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
 214
 215         PERF_SAMPLE_BRANCH_ANY          = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
 216         PERF_SAMPLE_BRANCH_ANY_CALL     = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
 217         PERF_SAMPLE_BRANCH_ANY_RETURN   = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
 218         PERF_SAMPLE_BRANCH_IND_CALL     = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
 219         PERF_SAMPLE_BRANCH_ABORT_TX     = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
 220         PERF_SAMPLE_BRANCH_IN_TX        = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
 221         PERF_SAMPLE_BRANCH_NO_TX        = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
 222         PERF_SAMPLE_BRANCH_COND         = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
 223
 224         PERF_SAMPLE_BRANCH_CALL_STACK   = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
 225         PERF_SAMPLE_BRANCH_IND_JUMP     = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
 226         PERF_SAMPLE_BRANCH_CALL         = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
 227
 228         PERF_SAMPLE_BRANCH_NO_FLAGS     = 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
 229         PERF_SAMPLE_BRANCH_NO_CYCLES    = 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
 230
 231         PERF_SAMPLE_BRANCH_TYPE_SAVE    =
 232                 1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
 233
 234         PERF_SAMPLE_BRANCH_HW_INDEX     = 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
 235
 236         PERF_SAMPLE_BRANCH_MAX          = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
 237 };
 238
 239 /*
 240  * Common flow change classification
 241  */
 242 enum {
 243         PERF_BR_UNKNOWN         = 0,    /* unknown */
 244         PERF_BR_COND            = 1,    /* conditional */
 245         PERF_BR_UNCOND          = 2,    /* unconditional  */
 246         PERF_BR_IND             = 3,    /* indirect */
 247         PERF_BR_CALL            = 4,    /* function call */
 248         PERF_BR_IND_CALL        = 5,    /* indirect function call */
 249         PERF_BR_RET             = 6,    /* function return */
 250         PERF_BR_SYSCALL         = 7,    /* syscall */
 251         PERF_BR_SYSRET          = 8,    /* syscall return */
 252         PERF_BR_COND_CALL       = 9,    /* conditional function call */
 253         PERF_BR_COND_RET        = 10,   /* conditional function return */
 254         PERF_BR_ERET            = 11,   /* exception return */
 255         PERF_BR_IRQ             = 12,   /* irq */
 256         PERF_BR_MAX,
 257 };
 258
 259 #define PERF_SAMPLE_BRANCH_PLM_ALL \
 260         (PERF_SAMPLE_BRANCH_USER|\
 261          PERF_SAMPLE_BRANCH_KERNEL|\
 262          PERF_SAMPLE_BRANCH_HV)
 263
 264 /*
 265  * Values to determine ABI of the registers dump.
 266  */
 267 enum perf_sample_regs_abi {
 268         PERF_SAMPLE_REGS_ABI_NONE       = 0,
 269         PERF_SAMPLE_REGS_ABI_32         = 1,
 270         PERF_SAMPLE_REGS_ABI_64         = 2,
 271 };
 272
 273 /*
 274  * Values for the memory transaction event qualifier, mostly for
 275  * abort events. Multiple bits can be set.
 276  */
 277 enum {
 278         PERF_TXN_ELISION        = (1 << 0), /* From elision */
 279         PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
 280         PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
 281         PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
 282         PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
 283         PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
 284         PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
 285         PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
 286
 287         PERF_TXN_MAX            = (1 << 8), /* non-ABI */
 288
 289         /* bits 32..63 are reserved for the abort code */
 290
 291         PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
 292         PERF_TXN_ABORT_SHIFT = 32,
 293 };
 294
 295 /*
 296  * The format of the data returned by read() on a perf event fd,
 297  * as specified by attr.read_format:
 298  *
 299  * struct read_format {
 300  *      { u64           value;
 301  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 302  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 303  *        { u64         id;           } && PERF_FORMAT_ID
 304  *      } && !PERF_FORMAT_GROUP
 305  *
 306  *      { u64           nr;
 307  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 308  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 309  *        { u64         value;
 310  *          { u64       id;           } && PERF_FORMAT_ID
 311  *        }             cntr[nr];
 312  *      } && PERF_FORMAT_GROUP
 313  * };
 314  */
 315 enum perf_event_read_format {
 316         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
 317         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
 318         PERF_FORMAT_ID                          = 1U << 2,
 319         PERF_FORMAT_GROUP                       = 1U << 3,
 320
 321         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
 322 };
 323
 324 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
 325 #define PERF_ATTR_SIZE_VER1     72      /* add: config2 */
 326 #define PERF_ATTR_SIZE_VER2     80      /* add: branch_sample_type */
 327 #define PERF_ATTR_SIZE_VER3     96      /* add: sample_regs_user */
 328                                         /* add: sample_stack_user */
 329 #define PERF_ATTR_SIZE_VER4     104     /* add: sample_regs_intr */
 330 #define PERF_ATTR_SIZE_VER5     112     /* add: aux_watermark */
 331 #define PERF_ATTR_SIZE_VER6     120     /* add: aux_sample_size */
 332 #define PERF_ATTR_SIZE_VER7     128     /* add: sig_data */
 333
 334 /*
 335  * Hardware event_id to monitor via a performance monitoring event:
 336  *
 337  * @sample_max_stack: Max number of frame pointers in a callchain,
 338  *                    should be < /proc/sys/kernel/perf_event_max_stack
 339  */
 340 struct perf_event_attr {
 341
 342         /*
 343          * Major type: hardware/software/tracepoint/etc.
 344          */
 345         __u32                   type;
 346
 347         /*
 348          * Size of the attr structure, for fwd/bwd compat.
 349          */
 350         __u32                   size;
 351
 352         /*
 353          * Type specific configuration information.
 354          */
 355         __u64                   config;
 356
 357         union {
 358                 __u64           sample_period;
 359                 __u64           sample_freq;
 360         };
 361
 362         __u64                   sample_type;
 363         __u64                   read_format;
 364
 365         __u64                   disabled       :  1, /* off by default        */
 366                                 inherit        :  1, /* children inherit it   */
 367                                 pinned         :  1, /* must always be on PMU */
 368                                 exclusive      :  1, /* only group on PMU     */
 369                                 exclude_user   :  1, /* don't count user      */
 370                                 exclude_kernel :  1, /* ditto kernel          */
 371                                 exclude_hv     :  1, /* ditto hypervisor      */
 372                                 exclude_idle   :  1, /* don't count when idle */
 373                                 mmap           :  1, /* include mmap data     */
 374                                 comm           :  1, /* include comm data     */
 375                                 freq           :  1, /* use freq, not period  */
 376                                 inherit_stat   :  1, /* per task counts       */
 377                                 enable_on_exec :  1, /* next exec enables     */
 378                                 task           :  1, /* trace fork/exit       */
 379                                 watermark      :  1, /* wakeup_watermark      */
 380                                 /*
 381                                  * precise_ip:
 382                                  *
 383                                  *  0 - SAMPLE_IP can have arbitrary skid
 384                                  *  1 - SAMPLE_IP must have constant skid
 385                                  *  2 - SAMPLE_IP requested to have 0 skid
 386                                  *  3 - SAMPLE_IP must have 0 skid
 387                                  *
 388                                  *  See also PERF_RECORD_MISC_EXACT_IP
 389                                  */
 390                                 precise_ip     :  2, /* skid constraint       */
 391                                 mmap_data      :  1, /* non-exec mmap data    */
 392                                 sample_id_all  :  1, /* sample_type all events */
 393
 394                                 exclude_host   :  1, /* don't count in host   */
 395                                 exclude_guest  :  1, /* don't count in guest  */
 396
 397                                 exclude_callchain_kernel : 1, /* exclude kernel callchains */
 398                                 exclude_callchain_user   : 1, /* exclude user callchains */
 399                                 mmap2          :  1, /* include mmap with inode data     */
 400                                 comm_exec      :  1, /* flag comm events that are due to an exec */
 401                                 use_clockid    :  1, /* use @clockid for time fields */
 402                                 context_switch :  1, /* context switch data */
 403                                 write_backward :  1, /* Write ring buffer from end to beginning */
 404                                 namespaces     :  1, /* include namespaces data */
 405                                 ksymbol        :  1, /* include ksymbol events */
 406                                 bpf_event      :  1, /* include bpf events */
 407                                 aux_output     :  1, /* generate AUX records instead of events */
 408                                 cgroup         :  1, /* include cgroup events */
 409                                 text_poke      :  1, /* include text poke events */
 410                                 build_id       :  1, /* use build id in mmap2 events */
 411                                 inherit_thread :  1, /* children only inherit if cloned with CLONE_THREAD */
 412                                 remove_on_exec :  1, /* event is removed from task on exec */
 413                                 sigtrap        :  1, /* send synchronous SIGTRAP on event */
 414                                 __reserved_1   : 26;
 415
 416         union {
 417                 __u32           wakeup_events;    /* wakeup every n events */
 418                 __u32           wakeup_watermark; /* bytes before wakeup   */
 419         };
 420
 421         __u32                   bp_type;
 422         union {
 423                 __u64           bp_addr;
 424                 __u64           kprobe_func; /* for perf_kprobe */
 425                 __u64           uprobe_path; /* for perf_uprobe */
 426                 __u64           config1; /* extension of config */
 427         };
 428         union {
 429                 __u64           bp_len;
 430                 __u64           kprobe_addr; /* when kprobe_func == NULL */
 431                 __u64           probe_offset; /* for perf_[k,u]probe */
 432                 __u64           config2; /* extension of config1 */
 433         };
 434         __u64   branch_sample_type; /* enum perf_branch_sample_type */
 435
 436         /*
 437          * Defines set of user regs to dump on samples.
 438          * See asm/perf_regs.h for details.
 439          */
 440         __u64   sample_regs_user;
 441
 442         /*
 443          * Defines size of the user stack to dump on samples.
 444          */
 445         __u32   sample_stack_user;
 446
 447         __s32   clockid;
 448         /*
 449          * Defines set of regs to dump for each sample
 450          * state captured on:
 451          *  - precise = 0: PMU interrupt
 452          *  - precise > 0: sampled instruction
 453          *
 454          * See asm/perf_regs.h for details.
 455          */
 456         __u64   sample_regs_intr;
 457
 458         /*
 459          * Wakeup watermark for AUX area
 460          */
 461         __u32   aux_watermark;
 462         __u16   sample_max_stack;
 463         __u16   __reserved_2;
 464         __u32   aux_sample_size;
 465         __u32   __reserved_3;
 466
 467         /*
 468          * User provided data if sigtrap=1, passed back to user via
 469          * siginfo_t::si_perf_data, e.g. to permit user to identify the event.
 470          * Note, siginfo_t::si_perf_data is long-sized, and sig_data will be
 471          * truncated accordingly on 32 bit architectures.
 472          */
 473         __u64   sig_data;
 474 };
 475
 476 /*
 477  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
 478  * to query bpf programs attached to the same perf tracepoint
 479  * as the given perf event.
 480  */
 481 struct perf_event_query_bpf {
 482         /*
 483          * The below ids array length
 484          */
 485         __u32   ids_len;
 486         /*
 487          * Set by the kernel to indicate the number of
 488          * available programs
 489          */
 490         __u32   prog_cnt;
 491         /*
 492          * User provided buffer to store program ids
 493          */
 494         __u32   ids[0];
 495 };
 496
 497 /*
 498  * Ioctls that can be done on a perf event fd:
 499  */
 500 #define PERF_EVENT_IOC_ENABLE                   _IO ('$', 0)
 501 #define PERF_EVENT_IOC_DISABLE                  _IO ('$', 1)
 502 #define PERF_EVENT_IOC_REFRESH                  _IO ('$', 2)
 503 #define PERF_EVENT_IOC_RESET                    _IO ('$', 3)
 504 #define PERF_EVENT_IOC_PERIOD                   _IOW('$', 4, __u64)
 505 #define PERF_EVENT_IOC_SET_OUTPUT               _IO ('$', 5)
 506 #define PERF_EVENT_IOC_SET_FILTER               _IOW('$', 6, char *)
 507 #define PERF_EVENT_IOC_ID                       _IOR('$', 7, __u64 *)
 508 #define PERF_EVENT_IOC_SET_BPF                  _IOW('$', 8, __u32)
 509 #define PERF_EVENT_IOC_PAUSE_OUTPUT             _IOW('$', 9, __u32)
 510 #define PERF_EVENT_IOC_QUERY_BPF                _IOWR('$', 10, struct perf_event_query_bpf *)
 511 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES        _IOW('$', 11, struct perf_event_attr *)
 512
 513 enum perf_event_ioc_flags {
 514         PERF_IOC_FLAG_GROUP             = 1U << 0,
 515 };
 516
 517 /*
 518  * Structure of the page that can be mapped via mmap
 519  */
 520 struct perf_event_mmap_page {
 521         __u32   version;                /* version number of this structure */
 522         __u32   compat_version;         /* lowest version this is compat with */
 523
 524         /*
 525          * Bits needed to read the hw events in user-space.
 526          *
 527          *   u32 seq, time_mult, time_shift, index, width;
 528          *   u64 count, enabled, running;
 529          *   u64 cyc, time_offset;
 530          *   s64 pmc = 0;
 531          *
 532          *   do {
 533          *     seq = pc->lock;
 534          *     barrier()
 535          *
 536          *     enabled = pc->time_enabled;
 537          *     running = pc->time_running;
 538          *
 539          *     if (pc->cap_usr_time && enabled != running) {
 540          *       cyc = rdtsc();
 541          *       time_offset = pc->time_offset;
 542          *       time_mult   = pc->time_mult;
 543          *       time_shift  = pc->time_shift;
 544          *     }
 545          *
 546          *     index = pc->index;
 547          *     count = pc->offset;
 548          *     if (pc->cap_user_rdpmc && index) {
 549          *       width = pc->pmc_width;
 550          *       pmc = rdpmc(index - 1);
 551          *     }
 552          *
 553          *     barrier();
 554          *   } while (pc->lock != seq);
 555          *
 556          * NOTE: for obvious reason this only works on self-monitoring
 557          *       processes.
 558          */
 559         __u32   lock;                   /* seqlock for synchronization */
 560         __u32   index;                  /* hardware event identifier */
 561         __s64   offset;                 /* add to hardware event value */
 562         __u64   time_enabled;           /* time event active */
 563         __u64   time_running;           /* time event on cpu */
 564         union {
 565                 __u64   capabilities;
 566                 struct {
 567                         __u64   cap_bit0                : 1, /* Always 0, deprecated, see commit 860f085b74e9 */
 568                                 cap_bit0_is_deprecated  : 1, /* Always 1, signals that bit 0 is zero */
 569
 570                                 cap_user_rdpmc          : 1, /* The RDPMC instruction can be used to read counts */
 571                                 cap_user_time           : 1, /* The time_{shift,mult,offset} fields are used */
 572                                 cap_user_time_zero      : 1, /* The time_zero field is used */
 573                                 cap_user_time_short     : 1, /* the time_{cycle,mask} fields are used */
 574                                 cap_____res             : 58;
 575                 };
 576         };
 577
 578         /*
 579          * If cap_user_rdpmc this field provides the bit-width of the value
 580          * read using the rdpmc() or equivalent instruction. This can be used
 581          * to sign extend the result like:
 582          *
 583          *   pmc <<= 64 - width;
 584          *   pmc >>= 64 - width; // signed shift right
 585          *   count += pmc;
 586          */
 587         __u16   pmc_width;
 588
 589         /*
 590          * If cap_usr_time the below fields can be used to compute the time
 591          * delta since time_enabled (in ns) using rdtsc or similar.
 592          *
 593          *   u64 quot, rem;
 594          *   u64 delta;
 595          *
 596          *   quot = (cyc >> time_shift);
 597          *   rem = cyc & (((u64)1 << time_shift) - 1);
 598          *   delta = time_offset + quot * time_mult +
 599          *              ((rem * time_mult) >> time_shift);
 600          *
 601          * Where time_offset,time_mult,time_shift and cyc are read in the
 602          * seqcount loop described above. This delta can then be added to
 603          * enabled and possible running (if index), improving the scaling:
 604          *
 605          *   enabled += delta;
 606          *   if (index)
 607          *     running += delta;
 608          *
 609          *   quot = count / running;
 610          *   rem  = count % running;
 611          *   count = quot * enabled + (rem * enabled) / running;
 612          */
 613         __u16   time_shift;
 614         __u32   time_mult;
 615         __u64   time_offset;
 616         /*
 617          * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
 618          * from sample timestamps.
 619          *
 620          *   time = timestamp - time_zero;
 621          *   quot = time / time_mult;
 622          *   rem  = time % time_mult;
 623          *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
 624          *
 625          * And vice versa:
 626          *
 627          *   quot = cyc >> time_shift;
 628          *   rem  = cyc & (((u64)1 << time_shift) - 1);
 629          *   timestamp = time_zero + quot * time_mult +
 630          *               ((rem * time_mult) >> time_shift);
 631          */
 632         __u64   time_zero;
 633
 634         __u32   size;                   /* Header size up to __reserved[] fields. */
 635         __u32   __reserved_1;
 636
 637         /*
 638          * If cap_usr_time_short, the hardware clock is less than 64bit wide
 639          * and we must compute the 'cyc' value, as used by cap_usr_time, as:
 640          *
 641          *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
 642          *
 643          * NOTE: this form is explicitly chosen such that cap_usr_time_short
 644          *       is a correction on top of cap_usr_time, and code that doesn't
 645          *       know about cap_usr_time_short still works under the assumption
 646          *       the counter doesn't wrap.
 647          */
 648         __u64   time_cycles;
 649         __u64   time_mask;
 650
 651                 /*
 652                  * Hole for extension of the self monitor capabilities
 653                  */
 654
 655         __u8    __reserved[116*8];      /* align to 1k. */
 656
 657         /*
 658          * Control data for the mmap() data buffer.
 659          *
 660          * User-space reading the @data_head value should issue an smp_rmb(),
 661          * after reading this value.
 662          *
 663          * When the mapping is PROT_WRITE the @data_tail value should be
 664          * written by userspace to reflect the last read data, after issueing
 665          * an smp_mb() to separate the data read from the ->data_tail store.
 666          * In this case the kernel will not over-write unread data.
 667          *
 668          * See perf_output_put_handle() for the data ordering.
 669          *
 670          * data_{offset,size} indicate the location and size of the perf record
 671          * buffer within the mmapped area.
 672          */
 673         __u64   data_head;              /* head in the data section */
 674         __u64   data_tail;              /* user-space written tail */
 675         __u64   data_offset;            /* where the buffer starts */
 676         __u64   data_size;              /* data buffer size */
 677
 678         /*
 679          * AUX area is defined by aux_{offset,size} fields that should be set
 680          * by the userspace, so that
 681          *
 682          *   aux_offset >= data_offset + data_size
 683          *
 684          * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
 685          *
 686          * Ring buffer pointers aux_{head,tail} have the same semantics as
 687          * data_{head,tail} and same ordering rules apply.
 688          */
 689         __u64   aux_head;
 690         __u64   aux_tail;
 691         __u64   aux_offset;
 692         __u64   aux_size;
 693 };
 694
 695 /*
 696  * The current state of perf_event_header::misc bits usage:
 697  * ('|' used bit, '-' unused bit)
 698  *
 699  *  012         CDEF
 700  *  |||---------||||
 701  *
 702  *  Where:
 703  *    0-2     CPUMODE_MASK
 704  *
 705  *    C       PROC_MAP_PARSE_TIMEOUT
 706  *    D       MMAP_DATA / COMM_EXEC / FORK_EXEC / SWITCH_OUT
 707  *    E       MMAP_BUILD_ID / EXACT_IP / SCHED_OUT_PREEMPT
 708  *    F       (reserved)
 709  */
 710
 711 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 712 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 713 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 714 #define PERF_RECORD_MISC_USER                   (2 << 0)
 715 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 716 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 717 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 718
 719 /*
 720  * Indicates that /proc/PID/maps parsing are truncated by time out.
 721  */
 722 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12)
 723 /*
 724  * Following PERF_RECORD_MISC_* are used on different
 725  * events, so can reuse the same bit position:
 726  *
 727  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
 728  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
 729  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
 730  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
 731  */
 732 #define PERF_RECORD_MISC_MMAP_DATA              (1 << 13)
 733 #define PERF_RECORD_MISC_COMM_EXEC              (1 << 13)
 734 #define PERF_RECORD_MISC_FORK_EXEC              (1 << 13)
 735 #define PERF_RECORD_MISC_SWITCH_OUT             (1 << 13)
 736 /*
 737  * These PERF_RECORD_MISC_* flags below are safely reused
 738  * for the following events:
 739  *
 740  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
 741  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
 742  *   PERF_RECORD_MISC_MMAP_BUILD_ID      - PERF_RECORD_MMAP2 event
 743  *
 744  *
 745  * PERF_RECORD_MISC_EXACT_IP:
 746  *   Indicates that the content of PERF_SAMPLE_IP points to
 747  *   the actual instruction that triggered the event. See also
 748  *   perf_event_attr::precise_ip.
 749  *
 750  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
 751  *   Indicates that thread was preempted in TASK_RUNNING state.
 752  *
 753  * PERF_RECORD_MISC_MMAP_BUILD_ID:
 754  *   Indicates that mmap2 event carries build id data.
 755  */
 756 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 757 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT     (1 << 14)
 758 #define PERF_RECORD_MISC_MMAP_BUILD_ID          (1 << 14)
 759 /*
 760  * Reserve the last bit to indicate some extended misc field
 761  */
 762 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 763
 764 struct perf_event_header {
 765         __u32   type;
 766         __u16   misc;
 767         __u16   size;
 768 };
 769
 770 struct perf_ns_link_info {
 771         __u64   dev;
 772         __u64   ino;
 773 };
 774
 775 enum {
 776         NET_NS_INDEX            = 0,
 777         UTS_NS_INDEX            = 1,
 778         IPC_NS_INDEX            = 2,
 779         PID_NS_INDEX            = 3,
 780         USER_NS_INDEX           = 4,
 781         MNT_NS_INDEX            = 5,
 782         CGROUP_NS_INDEX         = 6,
 783
 784         NR_NAMESPACES,          /* number of available namespaces */
 785 };
 786
 787 enum perf_event_type {
 788
 789         /*
 790          * If perf_event_attr.sample_id_all is set then all event types will
 791          * have the sample_type selected fields related to where/when
 792          * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
 793          * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
 794          * just after the perf_event_header and the fields already present for
 795          * the existing fields, i.e. at the end of the payload. That way a newer
 796          * perf.data file will be supported by older perf tools, with these new
 797          * optional fields being ignored.
 798          *
 799          * struct sample_id {
 800          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 801          *      { u64                   time;     } && PERF_SAMPLE_TIME
 802          *      { u64                   id;       } && PERF_SAMPLE_ID
 803          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 804          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 805          *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
 806          * } && perf_event_attr::sample_id_all
 807          *
 808          * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
 809          * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
 810          * relative to header.size.
 811          */
 812
 813         /*
 814          * The MMAP events record the PROT_EXEC mappings so that we can
 815          * correlate userspace IPs to code. They have the following structure:
 816          *
 817          * struct {
 818          *      struct perf_event_header        header;
 819          *
 820          *      u32                             pid, tid;
 821          *      u64                             addr;
 822          *      u64                             len;
 823          *      u64                             pgoff;
 824          *      char                            filename[];
 825          *      struct sample_id                sample_id;
 826          * };
 827          */
 828         PERF_RECORD_MMAP                        = 1,
 829
 830         /*
 831          * struct {
 832          *      struct perf_event_header        header;
 833          *      u64                             id;
 834          *      u64                             lost;
 835          *      struct sample_id                sample_id;
 836          * };
 837          */
 838         PERF_RECORD_LOST                        = 2,
 839
 840         /*
 841          * struct {
 842          *      struct perf_event_header        header;
 843          *
 844          *      u32                             pid, tid;
 845          *      char                            comm[];
 846          *      struct sample_id                sample_id;
 847          * };
 848          */
 849         PERF_RECORD_COMM                        = 3,
 850
 851         /*
 852          * struct {
 853          *      struct perf_event_header        header;
 854          *      u32                             pid, ppid;
 855          *      u32                             tid, ptid;
 856          *      u64                             time;
 857          *      struct sample_id                sample_id;
 858          * };
 859          */
 860         PERF_RECORD_EXIT                        = 4,
 861
 862         /*
 863          * struct {
 864          *      struct perf_event_header        header;
 865          *      u64                             time;
 866          *      u64                             id;
 867          *      u64                             stream_id;
 868          *      struct sample_id                sample_id;
 869          * };
 870          */
 871         PERF_RECORD_THROTTLE                    = 5,
 872         PERF_RECORD_UNTHROTTLE                  = 6,
 873
 874         /*
 875          * struct {
 876          *      struct perf_event_header        header;
 877          *      u32                             pid, ppid;
 878          *      u32                             tid, ptid;
 879          *      u64                             time;
 880          *      struct sample_id                sample_id;
 881          * };
 882          */
 883         PERF_RECORD_FORK                        = 7,
 884
 885         /*
 886          * struct {
 887          *      struct perf_event_header        header;
 888          *      u32                             pid, tid;
 889          *
 890          *      struct read_format              values;
 891          *      struct sample_id                sample_id;
 892          * };
 893          */
 894         PERF_RECORD_READ                        = 8,
 895
 896         /*
 897          * struct {
 898          *      struct perf_event_header        header;
 899          *
 900          *      #
 901          *      # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
 902          *      # The advantage of PERF_SAMPLE_IDENTIFIER is that its position
 903          *      # is fixed relative to header.
 904          *      #
 905          *
 906          *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
 907          *      { u64                   ip;       } && PERF_SAMPLE_IP
 908          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 909          *      { u64                   time;     } && PERF_SAMPLE_TIME
 910          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 911          *      { u64                   id;       } && PERF_SAMPLE_ID
 912          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 913          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 914          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 915          *
 916          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 917          *
 918          *      { u64                   nr,
 919          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 920          *
 921          *      #
 922          *      # The RAW record below is opaque data wrt the ABI
 923          *      #
 924          *      # That is, the ABI doesn't make any promises wrt to
 925          *      # the stability of its content, it may vary depending
 926          *      # on event, hardware, kernel version and phase of
 927          *      # the moon.
 928          *      #
 929          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 930          *      #
 931          *
 932          *      { u32                   size;
 933          *        char                  data[size];}&& PERF_SAMPLE_RAW
 934          *
 935          *      { u64                   nr;
 936          *        { u64 hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
 937          *        { u64 from, to, flags } lbr[nr];
 938          *      } && PERF_SAMPLE_BRANCH_STACK
 939          *
 940          *      { u64                   abi; # enum perf_sample_regs_abi
 941          *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
 942          *
 943          *      { u64                   size;
 944          *        char                  data[size];
 945          *        u64                   dyn_size; } && PERF_SAMPLE_STACK_USER
 946          *
 947          *      { union perf_sample_weight
 948          *       {
 949          *              u64             full; && PERF_SAMPLE_WEIGHT
 950          *      #if defined(__LITTLE_ENDIAN_BITFIELD)
 951          *              struct {
 952          *                      u32     var1_dw;
 953          *                      u16     var2_w;
 954          *                      u16     var3_w;
 955          *              } && PERF_SAMPLE_WEIGHT_STRUCT
 956          *      #elif defined(__BIG_ENDIAN_BITFIELD)
 957          *              struct {
 958          *                      u16     var3_w;
 959          *                      u16     var2_w;
 960          *                      u32     var1_dw;
 961          *              } && PERF_SAMPLE_WEIGHT_STRUCT
 962          *      #endif
 963          *       }
 964          *      }
 965          *      { u64                   data_src; } && PERF_SAMPLE_DATA_SRC
 966          *      { u64                   transaction; } && PERF_SAMPLE_TRANSACTION
 967          *      { u64                   abi; # enum perf_sample_regs_abi
 968          *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
 969          *      { u64                   phys_addr;} && PERF_SAMPLE_PHYS_ADDR
 970          *      { u64                   size;
 971          *        char                  data[size]; } && PERF_SAMPLE_AUX
 972          *      { u64                   data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
 973          *      { u64                   code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE
 974          * };
 975          */
 976         PERF_RECORD_SAMPLE                      = 9,
 977
 978         /*
 979          * The MMAP2 records are an augmented version of MMAP, they add
 980          * maj, min, ino numbers to be used to uniquely identify each mapping
 981          *
 982          * struct {
 983          *      struct perf_event_header        header;
 984          *
 985          *      u32                             pid, tid;
 986          *      u64                             addr;
 987          *      u64                             len;
 988          *      u64                             pgoff;
 989          *      union {
 990          *              struct {
 991          *                      u32             maj;
 992          *                      u32             min;
 993          *                      u64             ino;
 994          *                      u64             ino_generation;
 995          *              };
 996          *              struct {
 997          *                      u8              build_id_size;
 998          *                      u8              __reserved_1;
 999          *                      u16             __reserved_2;
1000          *                      u8              build_id[20];
1001          *              };
1002          *      };
1003          *      u32                             prot, flags;
1004          *      char                            filename[];
1005          *      struct sample_id                sample_id;
1006          * };
1007          */
1008         PERF_RECORD_MMAP2                       = 10,
1009
1010         /*
1011          * Records that new data landed in the AUX buffer part.
1012          *
1013          * struct {
1014          *      struct perf_event_header        header;
1015          *
1016          *      u64                             aux_offset;
1017          *      u64                             aux_size;
1018          *      u64                             flags;
1019          *      struct sample_id                sample_id;
1020          * };
1021          */
1022         PERF_RECORD_AUX                         = 11,
1023
1024         /*
1025          * Indicates that instruction trace has started
1026          *
1027          * struct {
1028          *      struct perf_event_header        header;
1029          *      u32                             pid;
1030          *      u32                             tid;
1031          *      struct sample_id                sample_id;
1032          * };
1033          */
1034         PERF_RECORD_ITRACE_START                = 12,
1035
1036         /*
1037          * Records the dropped/lost sample number.
1038          *
1039          * struct {
1040          *      struct perf_event_header        header;
1041          *
1042          *      u64                             lost;
1043          *      struct sample_id                sample_id;
1044          * };
1045          */
1046         PERF_RECORD_LOST_SAMPLES                = 13,
1047
1048         /*
1049          * Records a context switch in or out (flagged by
1050          * PERF_RECORD_MISC_SWITCH_OUT). See also
1051          * PERF_RECORD_SWITCH_CPU_WIDE.
1052          *
1053          * struct {
1054          *      struct perf_event_header        header;
1055          *      struct sample_id                sample_id;
1056          * };
1057          */
1058         PERF_RECORD_SWITCH                      = 14,
1059
1060         /*
1061          * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
1062          * next_prev_tid that are the next (switching out) or previous
1063          * (switching in) pid/tid.
1064          *
1065          * struct {
1066          *      struct perf_event_header        header;
1067          *      u32                             next_prev_pid;
1068          *      u32                             next_prev_tid;
1069          *      struct sample_id                sample_id;
1070          * };
1071          */
1072         PERF_RECORD_SWITCH_CPU_WIDE             = 15,
1073
1074         /*
1075          * struct {
1076          *      struct perf_event_header        header;
1077          *      u32                             pid;
1078          *      u32                             tid;
1079          *      u64                             nr_namespaces;
1080          *      { u64                           dev, inode; } [nr_namespaces];
1081          *      struct sample_id                sample_id;
1082          * };
1083          */
1084         PERF_RECORD_NAMESPACES                  = 16,
1085
1086         /*
1087          * Record ksymbol register/unregister events:
1088          *
1089          * struct {
1090          *      struct perf_event_header        header;
1091          *      u64                             addr;
1092          *      u32                             len;
1093          *      u16                             ksym_type;
1094          *      u16                             flags;
1095          *      char                            name[];
1096          *      struct sample_id                sample_id;
1097          * };
1098          */
1099         PERF_RECORD_KSYMBOL                     = 17,
1100
1101         /*
1102          * Record bpf events:
1103          *  enum perf_bpf_event_type {
1104          *      PERF_BPF_EVENT_UNKNOWN          = 0,
1105          *      PERF_BPF_EVENT_PROG_LOAD        = 1,
1106          *      PERF_BPF_EVENT_PROG_UNLOAD      = 2,
1107          *  };
1108          *
1109          * struct {
1110          *      struct perf_event_header        header;
1111          *      u16                             type;
1112          *      u16                             flags;
1113          *      u32                             id;
1114          *      u8                              tag[BPF_TAG_SIZE];
1115          *      struct sample_id                sample_id;
1116          * };
1117          */
1118         PERF_RECORD_BPF_EVENT                   = 18,
1119
1120         /*
1121          * struct {
1122          *      struct perf_event_header        header;
1123          *      u64                             id;
1124          *      char                            path[];
1125          *      struct sample_id                sample_id;
1126          * };
1127          */
1128         PERF_RECORD_CGROUP                      = 19,
1129
1130         /*
1131          * Records changes to kernel text i.e. self-modified code. 'old_len' is
1132          * the number of old bytes, 'new_len' is the number of new bytes. Either
1133          * 'old_len' or 'new_len' may be zero to indicate, for example, the
1134          * addition or removal of a trampoline. 'bytes' contains the old bytes
1135          * followed immediately by the new bytes.
1136          *
1137          * struct {
1138          *      struct perf_event_header        header;
1139          *      u64                             addr;
1140          *      u16                             old_len;
1141          *      u16                             new_len;
1142          *      u8                              bytes[];
1143          *      struct sample_id                sample_id;
1144          * };
1145          */
1146         PERF_RECORD_TEXT_POKE                   = 20,
1147
1148         /*
1149          * Data written to the AUX area by hardware due to aux_output, may need
1150          * to be matched to the event by an architecture-specific hardware ID.
1151          * This records the hardware ID, but requires sample_id to provide the
1152          * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT
1153          * records from multiple events.
1154          *
1155          * struct {
1156          *      struct perf_event_header        header;
1157          *      u64                             hw_id;
1158          *      struct sample_id                sample_id;
1159          * };
1160          */
1161         PERF_RECORD_AUX_OUTPUT_HW_ID            = 21,
1162
1163         PERF_RECORD_MAX,                        /* non-ABI */
1164 };
1165
1166 enum perf_record_ksymbol_type {
1167         PERF_RECORD_KSYMBOL_TYPE_UNKNOWN        = 0,
1168         PERF_RECORD_KSYMBOL_TYPE_BPF            = 1,
1169         /*
1170          * Out of line code such as kprobe-replaced instructions or optimized
1171          * kprobes or ftrace trampolines.
1172          */
1173         PERF_RECORD_KSYMBOL_TYPE_OOL            = 2,
1174         PERF_RECORD_KSYMBOL_TYPE_MAX            /* non-ABI */
1175 };
1176
1177 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER    (1 << 0)
1178
1179 enum perf_bpf_event_type {
1180         PERF_BPF_EVENT_UNKNOWN          = 0,
1181         PERF_BPF_EVENT_PROG_LOAD        = 1,
1182         PERF_BPF_EVENT_PROG_UNLOAD      = 2,
1183         PERF_BPF_EVENT_MAX,             /* non-ABI */
1184 };
1185
1186 #define PERF_MAX_STACK_DEPTH            127
1187 #define PERF_MAX_CONTEXTS_PER_STACK       8
1188
1189 enum perf_callchain_context {
1190         PERF_CONTEXT_HV                 = (__u64)-32,
1191         PERF_CONTEXT_KERNEL             = (__u64)-128,
1192         PERF_CONTEXT_USER               = (__u64)-512,
1193
1194         PERF_CONTEXT_GUEST              = (__u64)-2048,
1195         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
1196         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
1197
1198         PERF_CONTEXT_MAX                = (__u64)-4095,
1199 };
1200
1201 /**
1202  * PERF_RECORD_AUX::flags bits
1203  */
1204 #define PERF_AUX_FLAG_TRUNCATED                 0x01    /* record was truncated to fit */
1205 #define PERF_AUX_FLAG_OVERWRITE                 0x02    /* snapshot from overwrite mode */
1206 #define PERF_AUX_FLAG_PARTIAL                   0x04    /* record contains gaps */
1207 #define PERF_AUX_FLAG_COLLISION                 0x08    /* sample collided with another */
1208 #define PERF_AUX_FLAG_PMU_FORMAT_TYPE_MASK      0xff00  /* PMU specific trace format type */
1209
1210 /* CoreSight PMU AUX buffer formats */
1211 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_CORESIGHT        0x0000 /* Default for backward compatibility */
1212 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW              0x0100 /* Raw format of the source */
1213
1214 #define PERF_FLAG_FD_NO_GROUP           (1UL << 0)
1215 #define PERF_FLAG_FD_OUTPUT             (1UL << 1)
1216 #define PERF_FLAG_PID_CGROUP            (1UL << 2) /* pid=cgroup id, per-cpu mode only */
1217 #define PERF_FLAG_FD_CLOEXEC            (1UL << 3) /* O_CLOEXEC */
1218
1219 #if defined(__LITTLE_ENDIAN_BITFIELD)
1220 union perf_mem_data_src {
1221         __u64 val;
1222         struct {
1223                 __u64   mem_op:5,       /* type of opcode */
1224                         mem_lvl:14,     /* memory hierarchy level */
1225                         mem_snoop:5,    /* snoop mode */
1226                         mem_lock:2,     /* lock instr */
1227                         mem_dtlb:7,     /* tlb access */
1228                         mem_lvl_num:4,  /* memory hierarchy level number */
1229                         mem_remote:1,   /* remote */
1230                         mem_snoopx:2,   /* snoop mode, ext */
1231                         mem_blk:3,      /* access blocked */
1232                         mem_hops:3,     /* hop level */
1233                         mem_rsvd:18;
1234         };
1235 };
1236 #elif defined(__BIG_ENDIAN_BITFIELD)
1237 union perf_mem_data_src {
1238         __u64 val;
1239         struct {
1240                 __u64   mem_rsvd:18,
1241                         mem_hops:3,     /* hop level */
1242                         mem_blk:3,      /* access blocked */
1243                         mem_snoopx:2,   /* snoop mode, ext */
1244                         mem_remote:1,   /* remote */
1245                         mem_lvl_num:4,  /* memory hierarchy level number */
1246                         mem_dtlb:7,     /* tlb access */
1247                         mem_lock:2,     /* lock instr */
1248                         mem_snoop:5,    /* snoop mode */
1249                         mem_lvl:14,     /* memory hierarchy level */
1250                         mem_op:5;       /* type of opcode */
1251         };
1252 };
1253 #else
1254 #error "Unknown endianness"
1255 #endif
1256
1257 /* type of opcode (load/store/prefetch,code) */
1258 #define PERF_MEM_OP_NA          0x01 /* not available */
1259 #define PERF_MEM_OP_LOAD        0x02 /* load instruction */
1260 #define PERF_MEM_OP_STORE       0x04 /* store instruction */
1261 #define PERF_MEM_OP_PFETCH      0x08 /* prefetch */
1262 #define PERF_MEM_OP_EXEC        0x10 /* code (execution) */
1263 #define PERF_MEM_OP_SHIFT       0
1264
1265 /*
1266  * PERF_MEM_LVL_* namespace being depricated to some extent in the
1267  * favour of newer composite PERF_MEM_{LVLNUM_,REMOTE_,SNOOPX_} fields.
1268  * Supporting this namespace inorder to not break defined ABIs.
1269  *
1270  * memory hierarchy (memory level, hit or miss)
1271  */
1272 #define PERF_MEM_LVL_NA         0x01  /* not available */
1273 #define PERF_MEM_LVL_HIT        0x02  /* hit level */
1274 #define PERF_MEM_LVL_MISS       0x04  /* miss level  */
1275 #define PERF_MEM_LVL_L1         0x08  /* L1 */
1276 #define PERF_MEM_LVL_LFB        0x10  /* Line Fill Buffer */
1277 #define PERF_MEM_LVL_L2         0x20  /* L2 */
1278 #define PERF_MEM_LVL_L3         0x40  /* L3 */
1279 #define PERF_MEM_LVL_LOC_RAM    0x80  /* Local DRAM */
1280 #define PERF_MEM_LVL_REM_RAM1   0x100 /* Remote DRAM (1 hop) */
1281 #define PERF_MEM_LVL_REM_RAM2   0x200 /* Remote DRAM (2 hops) */
1282 #define PERF_MEM_LVL_REM_CCE1   0x400 /* Remote Cache (1 hop) */
1283 #define PERF_MEM_LVL_REM_CCE2   0x800 /* Remote Cache (2 hops) */
1284 #define PERF_MEM_LVL_IO         0x1000 /* I/O memory */
1285 #define PERF_MEM_LVL_UNC        0x2000 /* Uncached memory */
1286 #define PERF_MEM_LVL_SHIFT      5
1287
1288 #define PERF_MEM_REMOTE_REMOTE  0x01  /* Remote */
1289 #define PERF_MEM_REMOTE_SHIFT   37
1290
1291 #define PERF_MEM_LVLNUM_L1      0x01 /* L1 */
1292 #define PERF_MEM_LVLNUM_L2      0x02 /* L2 */
1293 #define PERF_MEM_LVLNUM_L3      0x03 /* L3 */
1294 #define PERF_MEM_LVLNUM_L4      0x04 /* L4 */
1295 /* 5-0xa available */
1296 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1297 #define PERF_MEM_LVLNUM_LFB     0x0c /* LFB */
1298 #define PERF_MEM_LVLNUM_RAM     0x0d /* RAM */
1299 #define PERF_MEM_LVLNUM_PMEM    0x0e /* PMEM */
1300 #define PERF_MEM_LVLNUM_NA      0x0f /* N/A */
1301
1302 #define PERF_MEM_LVLNUM_SHIFT   33
1303
1304 /* snoop mode */
1305 #define PERF_MEM_SNOOP_NA       0x01 /* not available */
1306 #define PERF_MEM_SNOOP_NONE     0x02 /* no snoop */
1307 #define PERF_MEM_SNOOP_HIT      0x04 /* snoop hit */
1308 #define PERF_MEM_SNOOP_MISS     0x08 /* snoop miss */
1309 #define PERF_MEM_SNOOP_HITM     0x10 /* snoop hit modified */
1310 #define PERF_MEM_SNOOP_SHIFT    19
1311
1312 #define PERF_MEM_SNOOPX_FWD     0x01 /* forward */
1313 /* 1 free */
1314 #define PERF_MEM_SNOOPX_SHIFT  38
1315
1316 /* locked instruction */
1317 #define PERF_MEM_LOCK_NA        0x01 /* not available */
1318 #define PERF_MEM_LOCK_LOCKED    0x02 /* locked transaction */
1319 #define PERF_MEM_LOCK_SHIFT     24
1320
1321 /* TLB access */
1322 #define PERF_MEM_TLB_NA         0x01 /* not available */
1323 #define PERF_MEM_TLB_HIT        0x02 /* hit level */
1324 #define PERF_MEM_TLB_MISS       0x04 /* miss level */
1325 #define PERF_MEM_TLB_L1         0x08 /* L1 */
1326 #define PERF_MEM_TLB_L2         0x10 /* L2 */
1327 #define PERF_MEM_TLB_WK         0x20 /* Hardware Walker*/
1328 #define PERF_MEM_TLB_OS         0x40 /* OS fault handler */
1329 #define PERF_MEM_TLB_SHIFT      26
1330
1331 /* Access blocked */
1332 #define PERF_MEM_BLK_NA         0x01 /* not available */
1333 #define PERF_MEM_BLK_DATA       0x02 /* data could not be forwarded */
1334 #define PERF_MEM_BLK_ADDR       0x04 /* address conflict */
1335 #define PERF_MEM_BLK_SHIFT      40
1336
1337 /* hop level */
1338 #define PERF_MEM_HOPS_0         0x01 /* remote core, same node */
1339 #define PERF_MEM_HOPS_1         0x02 /* remote node, same socket */
1340 #define PERF_MEM_HOPS_2         0x03 /* remote socket, same board */
1341 #define PERF_MEM_HOPS_3         0x04 /* remote board */
1342 /* 5-7 available */
1343 #define PERF_MEM_HOPS_SHIFT     43
1344
1345 #define PERF_MEM_S(a, s) \
1346         (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1347
1348 /*
1349  * single taken branch record layout:
1350  *
1351  *      from: source instruction (may not always be a branch insn)
1352  *        to: branch target
1353  *   mispred: branch target was mispredicted
1354  * predicted: branch target was predicted
1355  *
1356  * support for mispred, predicted is optional. In case it
1357  * is not supported mispred = predicted = 0.
1358  *
1359  *     in_tx: running in a hardware transaction
1360  *     abort: aborting a hardware transaction
1361  *    cycles: cycles from last branch (or 0 if not supported)
1362  *      type: branch type
1363  */
1364 struct perf_branch_entry {
1365         __u64   from;
1366         __u64   to;
1367         __u64   mispred:1,  /* target mispredicted */
1368                 predicted:1,/* target predicted */
1369                 in_tx:1,    /* in transaction */
1370                 abort:1,    /* transaction abort */
1371                 cycles:16,  /* cycle count to last branch */
1372                 type:4,     /* branch type */
1373                 reserved:40;
1374 };
1375
1376 union perf_sample_weight {
1377         __u64           full;
1378 #if defined(__LITTLE_ENDIAN_BITFIELD)
1379         struct {
1380                 __u32   var1_dw;
1381                 __u16   var2_w;
1382                 __u16   var3_w;
1383         };
1384 #elif defined(__BIG_ENDIAN_BITFIELD)
1385         struct {
1386                 __u16   var3_w;
1387                 __u16   var2_w;
1388                 __u32   var1_dw;
1389         };
1390 #else
1391 #error "Unknown endianness"
1392 #endif
1393 };
1394
1395 #endif /* _UAPI_LINUX_PERF_EVENT_H */